Why do fighting game AI suck?

[NeoAwesomeon]

So I was doing some training mode, and out of curiosity (and to see if the AI could show me some combos), I sparred with them a little bit, and they were the same input reading AI I've seen in so many other games.

Why is it that it's either input reading and frame one punishing you for a jab, or broken messes you're pretty sure never got tested and only try to use a combo that was patched out before the game even released? Is it just because we can't program them any other way?

This isn't just a KOF problem, just more of a trend I've noticed. How come no one has tried to be any better than this, focus on changing their reaction times to be closer to realistic, making sure they can function and be a decent challenge? Heck the only "fighting game" I know with decent AI is Smash Ultimate, it's not perfect but they strike a fair balance for the demographic they're targeting (Level 9 Kazuya is probably the best example I can give). My favorite fighting game as of writing this, Skullgirls, has super stupid AI who can barely combo and just spam projectiles unaware of how easy it is to close the gap.

My guess is devs think, "Bah, it's a fighting game! No one is going to play against the AI unless they want to fight the unfun masochistic final boss." and just never invest time (and money) into anything more than the bare minimum. Maybe it's "Fighting game AI were always ♥♥♥♥♥♥♥♥! If we change it now people will complain we aren't doing it justice!"

IDK, food for thought and hopefully an interesting prompt for discussion. If anyone has any good examples of AI done right, do share. I'd love to be proven wrong and see the exception to this rule I think I see.

[Fdynasty]

Originally posted by NeoAwesomeon:

So I was doing some training mode, and out of curiosity (and to see if the AI could show me some combos), I sparred with them a little bit, and they were the same input reading AI I've seen in so many other games.

Why is it that it's either input reading and frame one punishing you for a jab, or broken messes you're pretty sure never got tested and only try to use a combo that was patched out before the game even released? Is it just because we can't program them any other way?

This isn't just a KOF problem, just more of a trend I've noticed. How come no one has tried to be any better than this, focus on changing their reaction times to be closer to realistic, making sure they can function and be a decent challenge? Heck the only "fighting game" I know with decent AI is Smash Ultimate, it's not perfect but they strike a fair balance for the demographic they're targeting (Level 9 Kazuya is probably the best example I can give). My favorite fighting game as of writing this, Skullgirls, has super stupid AI who can barely combo and just spam projectiles unaware of how easy it is to close the gap.

My guess is devs think, "Bah, it's a fighting game! No one is going to play against the AI unless they want to fight the unfun masochistic final boss." and just never invest time (and money) into anything more than the bare minimum. Maybe it's "Fighting game AI were always ♥♥♥♥♥♥♥♥! If we change it now people will complain we aren't doing it justice!"

IDK, food for thought and hopefully an interesting prompt for discussion. If anyone has any good examples of AI done right, do share. I'd love to be proven wrong and see the exception to this rule I think I see.

why would you rely on computer to show you how to hit combos? there is already a section called mission where you can do trials and AI show you how to hit it in demo play

[diamond booger]

i think i remember there being an openai bot or something for sf4, it'd be nice if fighting game devs actually tried to implement meaningful AI. for example in chess you can learn things from computer AIs that humans wont do or havn't thought of

but i think a feature like this is probably too time consuming for them to code so they don't allocate budget to it

[SeeNoWeevil]

Fighting game AI has not improved at all over decades. It's the same thing copy and pasted for each new game. This is why (imo obviously) single player content in fighting games is pointless. If your AI doesn't behave even remotely like another human or just reacts instantly to the player's input, what is the point?? The whole concept of FGs breaks down.

Machine learning is almost certainly the solution but it just doesn't seem like anyone wants to invest in it.

[DrBobcat]

This is purely an issue that hasn't seen development because there hasn't been enough demand. I completely agree that AI in fighting games could be better, but it'll never happen so long as every word the devs see pertains to multiplayer. Just the reality of SP fighting game content and probably will be forever.

[NeoAwesomeon]

Originally posted by Fdynasty:

Originally posted by NeoAwesomeon:

So I was doing some training mode, and out of curiosity (and to see if the AI could show me some combos), I sparred with them a little bit, and they were the same input reading AI I've seen in so many other games.

Why is it that it's either input reading and frame one punishing you for a jab, or broken messes you're pretty sure never got tested and only try to use a combo that was patched out before the game even released? Is it just because we can't program them any other way?

This isn't just a KOF problem, just more of a trend I've noticed. How come no one has tried to be any better than this, focus on changing their reaction times to be closer to realistic, making sure they can function and be a decent challenge? Heck the only "fighting game" I know with decent AI is Smash Ultimate, it's not perfect but they strike a fair balance for the demographic they're targeting (Level 9 Kazuya is probably the best example I can give). My favorite fighting game as of writing this, Skullgirls, has super stupid AI who can barely combo and just spam projectiles unaware of how easy it is to close the gap.

My guess is devs think, "Bah, it's a fighting game! No one is going to play against the AI unless they want to fight the unfun masochistic final boss." and just never invest time (and money) into anything more than the bare minimum. Maybe it's "Fighting game AI were always ♥♥♥♥♥♥♥♥! If we change it now people will complain we aren't doing it justice!"

IDK, food for thought and hopefully an interesting prompt for discussion. If anyone has any good examples of AI done right, do share. I'd love to be proven wrong and see the exception to this rule I think I see.

why would you rely on computer to show you how to hit combos? there is already a section called mission where you can do trials and AI show you how to hit it in demo play

It varies from game to game, but most trails in fighting games do not give practical combos, just difficult ones. To use this game for example, the trials would lead me to beleive that Robert can only start combos with close D and nothing else. Athena can only combo from 6C, Krohnen can only combo from 2C, and so on so forth. Most of these combos require a lot of meter for not much reward or practicality.

I wanted to see how the AI acts and WHAT combos they go for, only to be hit with an AI that wont fight unless they can input read you. I did see the AI for Athena do a nice blockstring though and I'll probably try using something similar when I can. That's the kind of stuff I was looking for.

[Rook]

Why would you play against the AI and expect to learn from them?

[Amos]

Originally posted by GunMetalXIV:

Why would you play against the AI and expect to learn from them?

This mindset only exists because of the problem he mentioned before. You can learn from anything, and you would expect a new game to have lots of sort of things to mess around with. At least KOF 15 has endings and some unlockables, modern FGs don't even have those.
I agree with the OP. Fighting game developers don't (or probably can't?) invest any time into single player content. Granblue had the closest thing to actual content in recent time (a sidescrolling "RPG"), but it was lackluster.

Killer Instinct is the only fighting game I know of with good AI (Especially Shadow Brain clones) and decent single player content. (They've got a Shadow Lords campaign thing)

[aaaaaaaaaaaaa]

if u think it is easy build a game why dont urself make one?

[Cozmic Chaos]

At the very least they should set the ai to do trial combos.

[diamond booger]

Originally posted by GunMetalXIV:

Why would you play against the AI and expect to learn from them?

ever hear of openAI on dota 2? or deep blue in chess? learning AI is a thing and can be a great learning tool for players

[NeoAwesomeon]

Originally posted by GunMetalXIV:

Why would you play against the AI and expect to learn from them?

Learning can come from anywhere, and when my only options of things I can fight are people who've been playing this series for a decade or a moving punching bag, I as a new player would prefer the punching bag. Plus AI in games supposedly are created to fight as the developers intended them to, which in some scenarios means they have all the fundamentals down and that's what I need to get better at (though this usually isn't the case with fighting games as I am often reminded). AI players also aren't so good that I literally can't react, unlike fighting players online.
Also all of the online training rooms are already empty and I don't want to wake up my friends at midnight and be all "Yo dude can you spam fireballs and DPs at me for two hours I need to practice this jump-in."

Originally posted by 小聪明:

if u think it is easy build a game why dont urself make one?

I never said it was easy, I just said I doubt fighting game developers invest any time and money into developing the AI any more than they have to. As someone dabbling in game development as a hobby, I know it isn't something where they can just flip the "make AI good" switch. I am just wondering why off all aspects, fighting game AI hasn't improved from the old days of quarter eaters.
Quite frankly, this comment is incredibly dismissive.

Originally posted by 󠀡󠀡Cozmic:

At the very least they should set the ai to do trial combos.

Yeah, that's something that surprised me that they don't do. You'd think trial combos are the perfect example of how the devs want these characters to fight, but even the AI will use an auto combo. It's very strange to me.

[NeoAwesomeon]

Originally posted by diamond booger:

Originally posted by GunMetalXIV:

Why would you play against the AI and expect to learn from them?

ever hear of openAI on dota 2? or deep blue in chess? learning AI is a thing and can be a great learning tool for players

Honestly, I think having getting a learning AI would be the solution to this, but it leads me to a question I did think I'd have to ask. Did they already do this? AI learning tends to try to find the fastest/most effective route to success, and to me that'd explain a lot of frame 1 DPs. I think it's a good start, but a human hand would need to step in to fine tune these things afterwards. make it a 6-8 frame DP instead so it feels more like a human player that read you were going to jump in.
Not shooting down this idea though, watching AI play is how I learned to play Warcraft 3.

[NeoAwesomeon]

Originally posted by Amos:

Originally posted by GunMetalXIV:

Why would you play against the AI and expect to learn from them?

This mindset only exists because of the problem he mentioned before. You can learn from anything, and you would expect a new game to have lots of sort of things to mess around with. At least KOF 15 has endings and some unlockables, modern FGs don't even have those.
I agree with the OP. Fighting game developers don't (or probably can't?) invest any time into single player content. Granblue had the closest thing to actual content in recent time (a sidescrolling "RPG"), but it was lackluster.

Killer Instinct is the only fighting game I know of with good AI (Especially Shadow Brain clones) and decent single player content. (They've got a Shadow Lords campaign thing)

Okay, I can concede that Killer Instinct does have probably the best AI in the genre at the moment, but it can still be prone to BS (though most of that BS is Kyle difficulty because it learns how to fight you and breaks all of your combos).

[Fakeout the Breakout]

KOF's AI (or actually we should call it cheat logic) is rage-inducing but is better than nothing. Its useful for learning unsafe moves, to learn to take advantage of every little opportunity for optimal damage, and its satisfying to find and exploit the AI's weaknesses (like moves that cause the AI to do an unsafe roll, etc.).

A while ago I wrote fighting games with simple AIs for fun, and someday I hope to write a full algorithm driven mainly by frame data, collision boxes and distances. I don't think deep use of ML is needed unless you are trying to build an AI that comes up with its own fakes. At first we just need an intermediate AI that doesn't cheat.

For now I think SNK should provide a "no instant reaction" option where, even at the highest difficulty, the AI is forced to wait 200-300 milliseconds before reacting to the opponent's inputs. This will reduce its "challenge" substantially but will raise the usefulness of the AI as a training tool.

[Troispoint]

Developing a good AI is the most difficult thing you could try to do as a game dev. It's unfortunately not worth it given these games don't sell that well and SNK in particular ruined themselves financially by drawing the beautiful sprites of XIII. It's a potentially company ending risk to bet it all an advanced AI. Which we don't know if most people would care for anyway.